Year: 2014
Type: article
Abstract: Understanding what we mean by learning and how to accelerate learning is among the great challenges facing oncology. We are tasked to develop more sophisticated models of tumor biology and to learn how to apply tumor and host molecular profiles to direct treatment of patients with cancer. To accomplish that, we need to consider the clinical needs of the individual patient, as well as the opportunities to learn across populations of patients with cancer even as we subdivide anatomically defined cancers into multiple subpopulations categorized by tumor biology models driven by molecular test data. Biomarkers are the critical data elements that underlie how we understand cancer, develop new diagnostics and therapeutics, and improve patient lives. However, there are fundamental differences in the role biomarkers have depending on whether they are used to direct individual patient therapy or to create new learning. If learning is the understanding of how to apply knowledge to patient care, a learning health system distinguishes itself by the capacity to generate new learning, while delivering patient care.In its broadest sense, the term biomarker implies a measurement of a biological process whether as an assessment of health or disease.1 Although measurement may include various health data such as blood pressure or medical imaging results, in oncology the term is typically used for laboratory-based tests of abnormal germline or cancer cell genetics, gene and protein expression or modifications, and metabolites. Traditionally, clinicians have used singly measured abnormalities in a specific process to guide therapies (eg, overexpression of estrogen receptor protein to select endocrine therapy in breast cancer). More recently, the ability to perform deep whole-genome sequencing and analysis of DNA and RNA quickly and inexpensively has led to the ability to identify less common but potentially clinically important single abnormalities that may be used to determine future outcomes or serve as targets for novel therapies. Furthermore, advances in technology during the past decade have permitted interrogation of large numbers of abnormalities that can be combined into multiparameter signatures (genomics, transcriptomics, proteomics, and metabolomics), which lead to greater insights into the heterogeneous behavior of cancers and might provide guidance for clinical decision making in individual patients and a new classification of cancers based on laboratory pathology. These areas of science have been collectively referred to as omics.1Omics-based tests offer substantial promise in enabling the translation of laboratory-based observations to the subsequent realization of personalized or precision medicine. Tumor biomarker tests can be applied to 1 or more of the following possible uses that might guide clinical care more specifically: risk categorization, differential diagnosis (cancer versus benign tumor and the tissue of origin), prognosis (indication for therapy and level of intensity), prediction of whether a generic class or specific therapeutic agent will work, and monitoring patients during therapy to determine if a change in therapy is indicated and after therapy to detect disease recurrence. From the perspective of biomarkers and individual patients, precision or personalized medicine means getting the right treatment to the right patient at the right time. For this purpose, clinicians require a high degree of confidence that biomarker test results accurately support clinical decision making. Indeed, once established, highly validated tumor biomarker tests can be incorporated into digital clinical decision support systems (CDSSs) to guide treatment selection. The CDSSs are software systems that assess patient-specific data against a knowledge base using an algorithmic decision support engine.2 They are computerized versions of clinical decision support tools with the capability to personalize clinical support by drawing in patient-level data existing within the electronic health record. An example is logic that supports chemotherapy ordering by providing prespecified parameters on medication dosing and administration, which is triggered by patient-level information such as stage and biomarker data. Moreover, these systems can be configured to guide selection of the most appropriate tumor biomarker tests for each individual patient's situation. With a future that augurs an increasingly central role for biomarkers in directing cancer treatment, pathologists must appreciate the need for optimized and validated tumor biomarker tests that produce high-fidelity data and the growing use of robust CDSSs that depend on tumor biomarker test data. The CDSSs can be viewed as adding a point-of-care learning dimension to the health system.The word validated may be one of the most overused terms in the field. To clarify the issue, the Centers for Disease Control and Prevention convened the Evaluation of Genomic Applications in Practice and Prevention Working Group,3 which in 2009 recommended that genomics-based biomarker tests should be closely evaluated within the following 3 semantic categories: (1) analytic validity (the test has been shown to be accurate, reliable, and reproducible in the specimen type to be used in the clinical situation; test characteristics such as sensitivity, specificity, and reliability are confirmed); (2) clinical validity (the test separates a group into 2 or more distinct populations with different biological characteristics or clinical outcomes); and (3) clinical utility (the use of the test has been shown with high levels of evidence to improve clinical outcomes compared with not using the test).Whereas analytic and clinical validity transform data into knowledge, demonstration of clinical utility is the critical last step that allows for application of a tumor biomarker test in patient care. These terms and their significant distinctions must be fully understood as we enter the new world of next-generation sequencing and the development of omics-based tumor biomarker tests.Building on the Evaluation of Genomic Applications in Practice and Prevention definitions, the Institute of Medicine1 recommended that these 3 properties of a biomarker test should be achieved in a manner that ensures the trustworthiness of the biomarker test (Figure 1). The Institute of Medicine suggested that omics-based biomarker test generation should evolve through 3 separate steps, namely, discovery, test development, and determination of clinical utility, which emphasizes the importance of using separate discovery, analytic, and clinical validation data sets. They further insisted that computational models should be adequately described so that independent evaluation can occur, a recommendation that will pose some difficulty where proprietary computational models are involved. A Clinical Laboratory Improvement Amendments of 1988–certified laboratory should establish analytic and clinical validity.Perhaps most important, the Institute of Medicine endorsed rigorous demonstration of clinical utility either by conducting prospective clinical trials that test the test or by using archived specimens collected from previously conducted clinical trials to perform prospective retrospective studies.4 Regardless, they strongly recommended against the more commonly used strategy of testing the biomarker in samples of convenience in the absence of a locked-down test or a prospectively written analytic plan. Further demonstration of clinical utility with additional high levels of evidence can be generated in other clinical settings through faster observational and quasi-experimental studies, helping to accelerate learning.5The growing volume and complexity of tumor biomarker data create an increasing but unmet digital need. Electronic health records are required to store biomarker data in a manner that allows ready retrieval of such data for CDSSs and population health studies.As an example of the semantic specificity that will need to be addressed for categorization of biomarker data, one might consider what genomics means (Figure 2). Does it refer to a specific abnormality (mutation, copy number variation, translocation, etc) in the cancer DNA, and does it refer to such an abnormality regardless if it is detected by an allele-specific candidate gene approach or by total-genome sequencing? Does it imply a multiparameter gene expression signature that has been derived from an omics-based approach? Commercial manufacturers, and many academic laboratories, are already offering specific single-analyte tests, multiparameter RNA or protein signature tests, and analytic suites or panels of many different tests that provide a limited but fairly comprehensive profile of a tumor, and deep or next-generation sequencing results of the entire gene or transcriptome or both. Often, such test reports include detailed clinical guidance on the interpretation of these complex test results, and it is desirable to preserve this along with the test results.The amount of tumor biomarker data poised for clinical application is staggering. The human genome contains approximately 3 billion base pairs, representing a tetrabyte of digital data. That amount is doubled in light of the diploid nature of the human genome. For oncology, the volume of information is even greater if one includes both the static germline genome and the dynamic somatic genome in which variation over time can be associated with carcinogenesis or the development of heterogeneous cancer subpopulations within a patient. Therefore, the amount of information is yet again multiplied by additional samples obtained at clinically important time points such as first diagnosis, then at cancer recurrence, and later from additional sites of metastatic disease. Challenges related to data capture and representation have been discussed and go beyond the large amount of data generation for a single patient with cancer.6,7 Indeed, sophisticated bioinformatics analyses have become as important as the actual technical performance of the biomarker test. Pathologists must be aware of this deluge of data and its implications for the clinical laboratory of the future.The complexity does not stop here. One of the singular characteristics of biomarker data is that the value may increase with the passage of time. A DNA variant of unknown significance at the time of detection may be later identified as having a biological role in causation of disease. In contrast, a blood glucose level rapidly loses its clinical value in a matter of days, and the likelihood of new interpretive logic is considerably lower.Tumor and other specimen biobanks allow reanalysis of biological material as new technology becomes available and new biomarkers are identified. However, to do so and to provide clinically meaningful results will require information systems to track available samples and storage conditions, as well as the imperative to develop long-term archived tissue storage systems. Biomarker data will need to be available in machine-readable format to be searchable for clinical care and research purposes, annotated with clinical (phenotypic) information in electronic health records, and linked across time and health systems, including central cancer registries. Curation of tumor biomarker test data will require dedicated ontologies and dedicated information systems.8,9Data are transformed into knowledge when associated with other data in the context of a previously defined model that describes a relationship between these different types of data. Indeed, this is a working definition of an ontology. For single-analyte tests, the relationship is linear and transparent (eg, the long-recognized clinical utility of knowing the estrogen receptor status to determine the advisability of endocrine therapy for a patient with breast cancer). Single-analyte biomarker tests predominate in practice today, but what of the future?Biological systems are characterized by regulatory mechanisms such as feedback loops, alternative signal transduction pathways, and cross talk between molecular pathways. As a result, systems and computational biology will underlie the design of algorithm-based tumor biomarker tests. For example, examination of expression of thousands of genes led to the development of much smaller signatures such as the 21-gene recurrence score assay (Oncotype DX; Genomics Health Inc, Redwood City, California). This assay, which has been shown to have both analytic validity and clinical utility for guiding the use of chemotherapy in early-stage breast cancer, was ultimately generated by selection of approximately 300 candidate genes from published gene array data sets. These candidate genes were then winnowed down to the 21 genes that were ultimately incorporated into the test by annotating them against sets of archived tissues that had been collected in previously conducted prospective clinical trials. Pathologists will have an essential role in the design and implementation of tumor biomarker–driven cancer care. It will be imperative that the pathologists who are generating these data must understand that this future will be built on pathology laboratory reporting of tumor biomarker data in structured digital formats such as the College of American Pathologists electronic checklists. Once this is accomplished, biomarker data for individual patients can act as data inputs into computer applications, including CDSSs embedded in the electronic health record, that help guide treatment decision making. It will become essential that pathologists must insist that only tumor biomarker tests with proven analytic validity and clinical utility should be provided to the clinician for patient treatment decisions. The role of the pathologist in partnering with the clinician to select biomarker tests that have been validated and demonstrated to provide clinical value will become even more essential. With the creation of the College of American Pathologists Biomarkers Reporting Committee to standardize which tumor biomarkers are reported and which analytic methodology was used to assess test performance, the college has shown vision and leadership.A health care–based society might ask the following questions. How will we generate high levels of evidence for clinical utility as we begin to parse common cancers into increasingly uncommon subgroups based on multigene expression signatures or on single genetic abnormalities? How will we place these in the context of other omics abnormalities either known or perhaps unknown? Will a kidney cancer with the anaplastic lymphoma kinase (ALK) fusion oncogene be as sensitive to the ALK tyrosine kinase inhibitor crizotinib as lung cancer, or will there be further context-specific drivers of response and resistance? Can we afford to conduct prospective randomized trials for every drug in every genetic and biological subgroup? Is it possible to identify patients with outlier but exceptionally strong and durable responses to treatment and build n-of-1 (single subject) clinical trials10?Rapid-learning health systems (RLHSs) are strategies to aggregate large volumes of digital health data. They generate new evidence to support learning and drive changes in health care delivery.11The RLHSs feed on data sourced from routine health care operations. They reflect real-world experiences and more rapid accumulation of health data than is possible with clinical trials. Furthermore, RLHSs integrate societal and technological aspects to transform data into knowledge and knowledge into learning under the governance and for the purposes of a society.Proper use of an RLHS may permit documentation of the clinical utility of tumor biomarker tests through observational and quasi-experimental data collection, obviating or reducing the need for expensive and time-consuming prospective clinical trials.4 These large databases can be used to conduct comparative effectiveness research to determine if the use of tumor biomarker tests leads to improved outcomes. As opposed to the selection biases built into prospective trials, RLHSs provide population-based, patient-centered data, including not only improved survival but also symptom management and economic (improved health care value) or process (improved efficiency of care delivery) analyses. However, these types of studies will need to be performed carefully with sophisticated statistical adjustments, and until an outcome improvement by the use of the test compared with not using it can be shown, clinical utility has not been proven.The RLHSs can also enable data analysis of aggregated patient biomarker data to help develop systems biology models of cancer. Too often, heterogeneity of clinical behavior results from assigning cancer diagnosis by organ of origin. Rather, our increased understanding of cancer genetics and biology is suggesting that organ-denoted cancers are in fact many different diseases on the molecular level. These subtypes may often genetically and even phenotypically resemble cancers that have arisen in another organ more than they do other cancers from the same organ. Although not proven, the therapeutic implications are staggering. Instead of treating all breast cancers with one strategy and all colon cancers with another, it may be possible to more rapidly delineate different pathways by correlating cancer biomarker profiles with clinical behavior and outcomes. Therefore, we may in the future treat the pathway regardless of the tissue of origin.One implication of this new world is the necessity of gathering as much biomarker data as possible even if we do not yet understand how those data might be used. In other words, the accumulated biomarker data facilitate knowledge generation, which must of necessity precede the learning of how to use that knowledge to improve outcomes. However, accumulating such data requires collection of specimens, performance of the assays, and banking of the data. There are 2 major considerations to this approach. First, major objections exist to providing unproven test results to the patient and caregiver if we are not sure of the analytic validity or clinical utility.12 Second, if the test is not being performed for routine clinical care, who should for the biomarker testing and data These are with major The and of a next-generation has the need to the to and with the of patient of and of health has beyond in down the and increasing the of biomarker data However, the to tumor biomarker test analysis and the pressure on health care systems to new technology the of adding to health care Moreover, the of tumor biomarker tests with analytic validation and or evidence of clinical utility may in patient care or these 2 may regulatory and by and the regulatory and for tumor biomarker tests has been so have for and of both of these The that and tumor biomarker tests will improve patient care. In they that use of tumor biomarker tests will health care by of use of expensive in patients who are to of the from their was the of tests for patient These tests are permitted by of the and to facilitate the of various tests into the the and time-consuming process required for by the However, omics-based tests have under increased when the science the interpretation may not the for In that the and that a was in of the and This decision has a that the is on tests and will with and to the and use of on the use and reporting of tumor biomarker tests exist among often each the through its For example, the American College of and Genomics has a that all biomarker tests based on or genome sequencing methodology regardless of the of the test should be required to include genes that are associated with risk for significant The college that these genes should be whether or not those genes known relationship to the disease in As a of this patients and genetic risk information they not or were not to This recommendation to the of patient and the need for genetic However, it also supports to an degree the desirable to provide patients and to all data related to In a system such as health care, the use of data is a more complex than generating the to these are not However, the American of Clinical has the creation of a oncology will be provided digital tools clinical decision making based on and systems biology models that use tumor biomarker tests. In these will to clinical treatment and outcomes data into a large central Therefore, will serve as an RLHS that can be for understanding of the of omics-based data, which will be generated on most patients in the of data into learning a for the technical aspects of tumor biomarker tests in oncology. such as data and versus with and the between individual versus societal will require and across all in the health care These will require in such as the of on for example, the decision by the on of human genes and biomarker tests from other should be to understand the complexity of human biomarker testing societal implementation by such as the and and the of the for Health will will be by from of health care. by their with individual patients and with the through their have a in guiding this of society with They have an to to the rapidly knowledge base through new models of learning and knowledge was by the of on
Authors: Peter P. Yu, Mark A. Hoffman, Daniel F. Hayes
Language: English
Subfield: Cancer Research
Domain: Life Sciences
Open Access status: bronze
OpenAlex ID: W2051640697